⬆️ ⬇️

The first supercomputer DGX-1 based on Tesla V100 will be used in medicine

Scientists from the Center of Clinical Data Science will be the first to process data using a supercomputer for deep learning DGX-1 based on eight Tesla V100 graphics processors. V100 show a result of 960 teraflops with FP16 calculations thanks to Volta Tensor Core technology.





/ Flickr / Fritzchens Fritz / PD



The Tesla V100 data center platform was introduced in May 2017. It contains 21.1 billion transistors, built on a 12-nanometer FinFET process technology, and individual 640 Tensor cores are used to ensure the operation of neural networks, delivering 120 teraflops with in-depth training.

')

Nvidia has upgraded its NVLink bus — it now “develops” 300 Gbps, which is almost two times more than the previous implementation. This was made possible by increasing the number of contacts from four to six and expanding the bandwidth to 25 Gbit / s. The HBM2 3D memory module also received improvements - the bandwidth increased to 900 Gbps.



The Center of Clinical Data Science deals with automation and machine learning in the field of healthcare. They are working on a neural network that analyzes data on research on patients' diseases and then helps to make diagnoses faster.



“Doctors have to deal with a huge amount of information: laboratory tests, MRI, tomography, data on the health of family members and much more. Because of this, making decisions is incredibly difficult. The technology that will help doctors in diagnostics is able to optimize their work, ” said CCDS Executive Director Mark Michalski.



It is expected that in the future radiologists will have an assistant with artificial intelligence, who will help with diagnoses. Now doctors are studying pictures in the order in which they were taken. And artificial intelligence will be able to immediately determine which of them are the most problematic in order to send a specialist. Also, thanks to neural networks, assistants will be able to analyze images literally by pixels, and then compare them with other patient information and quickly diagnose them.



Even on the old equipment, researchers have developed algorithms for cardiac, ophthalmological, dermatological and psychiatric diagnostics. With the use of DGX-1 based on Volta video processors, these algorithms will become more accurate and more widely used.



Who else uses the GPU



More and more applications are supporting GPU computing, including frameworks for developing artificial intelligence. Therefore, many data centers that deal with deep learning work simultaneously with the GPU and CPU - the format is called heterogeneous computing. Thus, it is possible to take the best from both types of cores: the GPU copes with resource-intensive mathematical calculations, and the CPU “takes over” the work of the operating system and numerous simple operations.



According to a study by scientists at the University of California, compared with data centers with the same type of computing cores, heterogeneous ones have 21% higher performance and 23% higher energy efficiency.



Using the power of the GPU in their systems of artificial intelligence and machine learning company Facebook. The corresponding laboratory within the company develops neural networks for solving specific problems.



According to experts of the region, the volume of data collected by companies is increasing. Therefore, GPUs are beginning to be used not only to work with demanding computing for learning neural networks. But also to work with databases.



For example, Nike uses servers with GPU and MapD software to analyze sales history and predict demand in selected regions. Another MapD client, Verizon, uses systems with a GPU to analyze server logs that track mobile phones.



They offer work with GPU-accelerated servers and cloud providers. Including IT-GRAD company. The solution allows you to experiment with analytical or requiring visual support projects. GPU systems enable organizations to quickly analyze large data arrays and in some particular situations can replace entire server clusters.



PS Some materials on the topic from our blog:



Source: https://habr.com/ru/post/338212/



All Articles